摘要
目的探讨钆塞酸二钠(Gd-EOB-DTPA)增强MRI深度学习模型术前预测肝细胞癌(HCC)血管包绕肿瘤细胞巢(VETC)的价值。方法本研究为病例-对照研究, 回顾性收集2016年6月至2023年3月在接受Gd-EOB-DTPA增强MRI检查并经手术病理证实为HCC的420例患者资料, 来自苏州大学附属第一医院的305例患者为训练集, 来自南通大学附属南通第三医院的115例患者为外部验证集。根据术后病理将患者分为VETC阳性组和VETC阴性组, 训练集中VETC阳性组161例、阴性组144例, 外部验证集中VETC阳性组55例、阴性组60例。使用ITK-SNAP软件手动勾画动脉期、门静脉期和肝胆期的肿瘤感兴趣区体积, 利用预训练的Vgg19、Densenet121和Vision Transformer(ViT)模型进行迁移学习, 并分别提取每张图像的深度学习特征。利用FAE软件对特征数据进行处理, 并采用logistic回归构建动脉期、门静脉期、肝胆期和三期联合深度学习模型共12种, 筛选出最优深度学习模型。对临床特征进行单因素和多因素分析并筛选出独立的预测指标构建临床模型。联合临床独立预测指标及最优深度学习特征构建临床-深度学习模型。使用受试者操作特征曲线对模型预测VETC阳性HCC的效能进行评价。绘制校准曲线和决策曲线评估模型的预测准确性和临床实用性。结果在外部验证集, Vgg19模型在动脉期、门静脉期、肝胆期及三期联合的曲线下面积(AUC)为0.799、0.756、0.789、0.821, 均高于Densenet121的AUC(0.544、0.581、0.544、0.583)和ViT的AUC(0.740、0.752、0.785、0.767)。其中三期联合Vgg19模型预测效能最优, AUC为0.821(95%CI 0.746~0.897)。多因素logistic回归显示甲胎蛋白(OR=1.826, 95%CI 1.069~3.120, P=0.028)和肿瘤长径(OR=1.329, 95%CI 1.206~1.466, P<0.001)是VETC阳性HCC的独立预测因子。基于这2个临床指标构建的临床模型和临床-深度学习模型的AUC分别为0.789(95%CI 0.703~0.859)和0.825(95%CI 0.749~0.900)。校准曲线提示三期联合Vgg19模型预测概率与实际预测概率之间具有较高的准确性, 外部验证集的决策曲线提示, 对比临床模型, 三期联合Vgg19模型和临床-深度学习模型预测VETC具有更好的净获益。结论基于Gd-EOB-DTPA增强MRI的深度学习模型可用于术前预测HCC的VETC模式, 其中三期联合Vgg19模型和临床-深度学习模型具有较高的预测价值。
Objective To explore the value of the deep learning model based on gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid(Gd-EOB-DTPA)enhanced MRI in preoperatively predicting vessels completely encapsulating tumor clusters(VETC)in hepatocellular carcinoma(HCC).Methods This study adopted a case-control design to retrospectively analyze 420 patients with HCC confirmed by postoperative pathology who underwent Gd-EOB-DTPA enhanced MRI between June 2016 and March 2023.A total of 420 patients were divided into a training set(n=305)from the First Affiliated Hospital of Soochow University and an external validation set(n=115)from Affiliated Nantong Hospital 3 of Nantong University.Based on postoperative pathological findings,patients were stratified into VETC-positive and VETC-negative groups.The training set comprised 161 VETC-positive cases and 144 VETC-negative cases,while the external validation set included 55 VETC-positive cases and 60 VETC-negative cases.Tumor regions of interest in arterial,portal venous,and hepatobiliary phases were manually delineated using ITK-SNAP software.Pre-trained Vgg19,Densenet121,and Vision Transformer(ViT)models were employed for transfer learning,extracting deep learning features from each image.Feature data were processed using FAE software,and 12 logistic regression models(arterial phase,portal venous phase,hepatobiliary phase,and combined three-phase models)were constructed to select the optimal deep learning model.Independent predictors in clinical characteristics were identified through univariate and multivariate logistic analyses to establish a clinical model for predicting VETC pattern.Subsequently,a clinical-deep learning fusion model was developed by integrating these clinical predictors with the optimal deep learning features.Model performance in predicting VETC-positive HCC was evaluated using receiver operating characteristic curves,calibration curves,and decision curve analysis(DCA).Results In the external validation set,the area under the curve(AUC)of the Vgg19 model in the arterial phase,portal venous phase,hepatobiliary phase,and combined three-phase,respectively were 0.799,0.756,0.789,0.821,which were higher than those of Densenet121(AUC:0.544,0.581,0.544,0.583)and ViT(AUC:0.740,0.752,0.785,0.767)model.The three-phase combined Vgg19 model achieved the highest AUC of 0.821(95%CI 0.746-0.897).Multivariate logistic regression identified alpha-fetoprotein level(0R=1.826,95%CI 1.069-3.120,P=0.028)and tumor diameter(0R=1.329,95%CI 1.206-1.466,P<0.001)as independent predictors of VETC-positive HCC,forming the clinical model with an AUC of 0.789(95%CI 0.703-0.859).The clinical-deep learning fusion model further achieved the AUC of 0.825(95%CI 0.749-0.900).Calibration curves confirmed high concordance between predicted and actual probabilities for the three-phase Vgg19 model,while DCA revealed greater net clinical benefit for the combined Vgg19 and fusion models compared with the clinical model alone.Conclusions The deep learning model based on Gd-EOB-DTPA enhanced MRI can be used to predict VETC of HCC preoperatively,among which the three-phase combined Vgg19 model and the clinical-deep learning model provide high predictivevalue.
作者
王锦晶
诗涔
范艳芬
吴茜
张涛
张继云
顾文豪
王希明
胡春洪
郁义星
Wang Jinjing;Shi Cen;Fan Yanfen;Wu Qian;Zhang Tao;Zhang Jiyun;Gu Wenhao;Wang Ximing;Hu Chunhong;Yu Yixing(Department of Radiology,the First Affiliated Hospital of Soochow University,Institute of Imaging Medicine,Soochow University,Suzhou 215006,China;Department of Radiology,Affiliated Nantong Hospital 3 of Nantong University,Nantong 226021,China;Department of Radiology,the First People's Hospital of Taicang City,Suzhou 215400,China)
出处
《中华放射学杂志》
北大核心
2025年第6期657-664,共8页
Chinese Journal of Radiology
基金
中国博士后科学基金面上项目(2024M752334)
苏州科技计划项目(SKY2023146)
苏州市基础研究试点项目(SSD2024083)。
关键词
癌
肝细胞
磁共振成像
血管包绕肿瘤细胞巢
钆塞酸二钠
深度学习
Carcinoma,hepatocellular
Magneticresonance imaging
Vessels completelyencapsulating tumor clusters
Gadoliniumethoxybenzyl diethylenetriamine pentaaceticacid
Deep learning
